#2489 kernel benches: host-timed (graph + barrier-event profiling hung on an A770 under xe); native_grouped_parity --bench2

说明

## Title
kernel benches: time on the host (graph + barrier-event profiling hung on an A770 under xe); native_grouped_parity --bench2

## Summary
On an Arc A770 under xe (compute runtime 26.05, oneAPI 2026.1), `dense_bench` and `native_grouped_parity --bench` never returned. Both replay a graph timed by the profiling info of `ext_oneapi_submit_barrier` events, and the process span in the kernel for 25 minutes. The engine's own graphs are unaffected.

## What changed
- `sycl/src/kernels/dense_bench.cpp`: times each graph launch on the host, with `queue::wait`. Graph launch cost is now included in the figure.
- `sycl/src/kernels/native_grouped_parity.cpp`: new `--bench2 <gu type> <down type> [groups entries]`. It times one format pair at the Coder's decode shape (2560 × 640, 4-token window × 10), 48 layers with their own experts, as direct launches. It reports µs per call, the gate/up and down phases separately, and the GB/s of the experts read. The existing modes are unchanged.

## Extra Notes
Timing only; the checks and hashes are unchanged.

🤖 Generated with [Claude Code](https://claude.com/claude-code)